CEMP: Clean Energy Materials Platform

• CEMP is an open-access digital ecosystem unifying high-throughput simulation, AI predictions, and curated databases to drive rapid clean-energy materials innovation.
• The platform integrates multi-scale computations, online validation, and closed-loop workflows, enabling coordinated screening across ionic liquids, polymers, crystals, and batteries.
• CEMP employs dedicated modules and ML models combined with quantum and molecular dynamics simulations to optimize material properties and support cross-domain applications.

Searching arXiv for recent and related work on Clean Energy Materials Platform and materials acceleration platforms. The Clean Energy Materials Platform (CEMP) is an open-access digital ecosystem for discovering and optimizing clean-energy materials by unifying high-throughput online calculation, multi-scale machine learning, and a heterogeneous materials database within a closed-loop workflow from data acquisition to prediction, simulation-based validation, and database expansion (Wang et al., 6 Jul 2025). In its current formulation, CEMP is tailored to four material classes—small molecules, polymers, ionic liquids, and crystals—and emphasizes clean-energy-relevant properties that are often weakly represented in crystal-centric platforms, including ionic conductivity, electrochemical window, mechanical properties, electrode voltage, and battery discharge behavior (Wang et al., 6 Jul 2025). More broadly, the platform belongs to the same methodological lineage as self-driving laboratories and materials acceleration platforms, in which automated experimentation, characterization, databases, and AI-guided decision-making are treated as a single integrated infrastructure rather than as disconnected tools (MacLeod et al., 2019, Wang et al., 2023).

1. Conceptual scope and design rationale

CEMP is designed to address three gaps in the contemporary materials-informatics landscape: limited chemical-space coverage beyond inorganic crystals, separation between databases and executable computation, and the absence of a genuinely closed-loop framework for clean-energy materials research (Wang et al., 6 Jul 2025). Its stated scope explicitly extends beyond electrodes and crystalline solids to small molecules, ionic liquids, and polymers, thereby covering solvents, salts, monomers, polymer electrolytes, binders, separators, and functional solids within a single framework (Wang et al., 6 Jul 2025).

This breadth is consequential because many clean-energy problems are intrinsically cross-domain. Battery design, for example, depends simultaneously on crystalline electrode hosts, molecular or ionic liquid electrolytes, polymeric separators, and device-level operating conditions. CEMP therefore frames discovery not as isolated prediction of one class of material, but as coordinated screening across multiple classes and scales (Wang et al., 6 Jul 2025). A plausible implication is that the platform is intended to support co-optimization of materials ensembles rather than only ranking individual compounds.

The platform’s closed-loop logic is explicit: data acquisition $\rightarrow$ model training $\rightarrow$ model prediction $\rightarrow$ validation via online simulation $\rightarrow$ database expansion $\rightarrow$ improved models (Wang et al., 6 Jul 2025). This workflow closely parallels the general MAP architecture proposed for heterogeneous CO$_2$ photo(thermal)catalysis, where initialization, AI control, experimentation, characterization, and database feedback are treated as coupled stages in a self-driving discovery cycle (Wang et al., 2023). It also resembles the modular architecture of the “Ada” self-driving laboratory, which used ChemOS and Phoenics to propose experiments, execute them robotically, analyze results, and update the optimization model in parallel (MacLeod et al., 2019).

2. System architecture and platform modules

CEMP is implemented on a Django/Python backend with standard web technologies on the front end, and is organized into five core modules: an Autocompute module, an Ionic Liquid module, a Polymer module, a Crystal module, and a Battery Management System (BMS) module (Wang et al., 6 Jul 2025). At the interaction level, users access the platform through web forms, choose a module, and either query the database, submit structures for prediction, or launch online calculations by uploading a structured spreadsheet with SMILES and parameters (Wang et al., 6 Jul 2025).

The architecture couples three functions that are often separated in other infrastructures. First, it serves as a database containing experimental measurements, quantum-chemical calculations, and AI-predicted properties. Second, it exposes machine-learning predictors that operate directly on user-submitted molecular or crystal representations. Third, it provides online execution of quantum chemistry and molecular dynamics workflows, allowing model outputs to be checked by physics-based simulation without leaving the platform (Wang et al., 6 Jul 2025). This integration distinguishes it from static repositories and from stand-alone prediction services.

The data flow is sequential but recursive. Experimental data from literature and databases are curated and standardized; QC and MD data are generated through Autocompute; ML models are trained on these accumulated records; users invoke models for property prediction or screening; promising candidates are then validated through ORCA or GROMACS jobs; and the resulting outputs are stored back into the database (Wang et al., 6 Jul 2025). This architecture instantiates, in web-platform form, the “database $\leftrightarrow$ ML $\leftrightarrow$ automated validation” philosophy advocated in MAP design studies (Wang et al., 2023).

CEMP’s web interface operationalizes this architecture through dedicated prediction pages for ionic liquids, polymers, crystals, and BMS tasks, together with QC and MD submission interfaces and JSmol-based structure viewers (Wang et al., 6 Jul 2025). The platform also includes an internal computation log to prevent redundant calculations, which is functionally analogous to orchestration mechanisms in autonomous labs that schedule tasks and reuse prior results when possible (Wang et al., 6 Jul 2025, MacLeod et al., 2019).

3. Data model, material classes, and property coverage

CEMP’s database is FAIR-compliant and integrates four material classes with heterogeneous representations: SMILES-based small molecules, SMILES-defined ionic liquid cation–anion pairs, SMILES-based polymer repeating units, and CIF-based crystals converted into graph structures via pymatgen (Wang et al., 6 Jul 2025). The database hosts about 376,000 entries, including about 6,000 experimental records, about 50,000 quantum-chemical calculations, and about 320,000 AI-predicted properties (Wang et al., 6 Jul 2025). By sub-database, it contains about 20,000 molecular QC records, about 8,000 ionic-liquid QC entries plus about 2,000 experimental ionic-liquid entries and about 100,000 ML predictions, about 10,000 polymer QC entries plus about 4,000 experimental polymer records and about 210,000 ML-predicted polymer records, and about 10,000 crystal entries imported from Materials Project (Wang et al., 6 Jul 2025).

Standardized QC properties include total energy, HOMO energy, LUMO energy, HOMO–LUMO gap, enthalpy $H$, entropy $S$, Gibbs free energy $\rightarrow$0, and dipole moment (Wang et al., 6 Jul 2025). On top of these, the platform emphasizes 12 critical modeled properties spanning ionic liquids, polymers, crystals, and batteries: melting point, electrochemical window, ionic conductivity, glass transition temperature, tensile strength, Young’s modulus, dielectric constant, average voltage, specific capacity, specific energy, C-rate discharge performance, and electrochemical redox properties (Wang et al., 6 Jul 2025). The polymer database extends beyond this set to 26 properties including decomposition temperature, water contact angle, refractive index, thermal conductivity, elongation at break, gas permeabilities, ion exchange capacity, swelling, water uptake, critical solution temperatures, bandgap, and limiting oxygen index (Wang et al., 6 Jul 2025).

A central feature of CEMP is that all records carry material identifiers, source labels distinguishing QC, experiment, and ML, and method-level metadata such as level of theory, software, force field, MD parameters, or model type and performance metrics (Wang et al., 6 Jul 2025). Experimental aggregation uses outlier removal by the $\rightarrow$1 rule and robust averaging by a 50% trimmed mean (Wang et al., 6 Jul 2025). This standardization is essential because clean-energy datasets are intrinsically heterogeneous and often condition-sensitive, a point also emphasized in MAP design analyses for catalysis and scale-up (Wang et al., 2023).

The platform’s data model is well suited to host not only native CEMP calculations but also structured external studies. For example, Janus chromium monolayers with recorded dipole moments, face-resolved work-function asymmetry, HSE06 band gaps, ferromagnetic ordering energies, and site-resolved HER descriptors constitute the kind of multi-property 2D entries that can populate a clean-energy platform with coupled electronic, catalytic, and interfacial metadata (Li et al., 2022). Likewise, T-carbon can be represented as a crystal entry with structural prototype, band gap, thermoelectric descriptors, hydrogen-storage capacity, ion-diffusion barriers, and application tags for batteries and hydrogen storage (Qin et al., 2019). Ce-doped graphene ORR catalysts supply another example of platform-compatible entries in which structure archetypes, formation energies, CHE-derived overpotentials, and pathway selectivity are explicitly defined (Lucchetti et al., 9 May 2025). Emulsion-templated inorganic/inorganic composites add a process-centric class of records where processing variables, microstructural descriptors, and mechanical performance are jointly relevant to energy applications (Jiang et al., 2024).

4. High-throughput online calculation infrastructure

The Autocompute module is the computational backbone of CEMP. It accepts spreadsheet uploads specifying molecule name, SMILES string, MD variables such as number of molecules, temperature, and simulation time, and task type such as QC geometry optimization, dimer binding energy, redox potential, or MD diffusion and cohesion (Wang et al., 6 Jul 2025). Preprocessing converts SMILES to 3D geometries using OpenBabel by generating 50 conformers, minimizing them with UFF, and selecting the lowest-energy conformer; formal charge and multiplicity are assigned automatically; and QC or MD input files are generated accordingly (Wang et al., 6 Jul 2025).

Scheduling dispatches QC jobs to CPU nodes running ORCA and MD jobs to GPU nodes running GROMACS, while monitoring server load and currently allowing up to three concurrent tasks (Wang et al., 6 Jul 2025). The QC branch implements a standard protocol consisting of geometry optimization at B3LYP/def2-TZVP plus D3, frequency calculation at the same level to confirm the absence of imaginary frequencies, and a single-point energy at $\rightarrow$2B97M-V/def2-TZVP (Wang et al., 6 Jul 2025). Error handling detects convergence failures and imaginary frequencies, adjusts starting wavefunctions or keywords, and retries jobs (Wang et al., 6 Jul 2025).

The MD branch uses GAFF for organic species, OPC3 water, Merz parameters for metal ions, and RESP charges scaled by 0.7 to mimic condensed-phase polarization (Wang et al., 6 Jul 2025). Its simulation protocol employs a leap-frog integrator with a 2 fs timestep, 3D periodic boundary conditions, 1.2 nm cutoffs for van der Waals and short-range electrostatics, SPME for long-range electrostatics, a velocity-rescale thermostat with $\rightarrow$3 ps, a Berendsen barostat with $\rightarrow$4 ps, and LINCS bond constraints, preceded by energy minimization, 5 ns NPT equilibration, 5 ns annealing NVT from 400 K to target temperature, and production NVT runs for user-specified time (Wang et al., 6 Jul 2025).

Output includes QC electronic properties, thermodynamic properties, optimized structures, and full ORCA files, as well as MD diffusion coefficients, density, viscosity, ionic conductivity, RDFs, coordination numbers, mean residence time, trajectories, topology files, logs, and plots (Wang et al., 6 Jul 2025). Some properties are given explicitly through standard relations, including the diffusion coefficient

$\rightarrow$5

and redox energies

$\rightarrow$6

(Wang et al., 6 Jul 2025).

This online-computation layer gives CEMP a capability that many materials databases lack: immediate validation of predicted candidates through structured, platform-native QC or MD. In the language of broader MAP design, this is the “physics-based validation” component required to prevent purely statistical pipelines from becoming detached from mechanistic reality (Wang et al., 2023). It also echoes autonomous-laboratory practice, where proposal, execution, and analysis are treated as parts of one orchestrated loop rather than as separate stages (MacLeod et al., 2019).

5. Machine learning models and predictive performance

CEMP deploys 12 ML models across its modules, with model family selected by data type and task (Wang et al., 6 Jul 2025). For molecules, ionic liquids, and polymers, inputs are SMILES strings featurized as 2048-bit Morgan fingerprints with radius 2 generated using RDKit (Wang et al., 6 Jul 2025). For crystals, CIF files are converted into graphs whose nodes are atoms and whose edges connect neighbors within 5 Å, with distances expanded by Gaussian basis functions (Wang et al., 6 Jul 2025). For battery C-rate prediction, the model consumes a feature vector encoding cathode type, electrolyte composition, electrolyte thickness, temperature, and related variables (Wang et al., 6 Jul 2025).

The deployed model families are XGBoost, multilayer perceptrons, graph neural networks including GCN and GAT, a MOCO-enhanced GAT representation learner for crystals, and a Transformer for discharge-curve prediction (Wang et al., 6 Jul 2025). The crystal workflow is especially notable because it combines self-supervised contrastive pretraining with regression fine-tuning: MOCO uses query and momentum encoders, random omission of 1–5% of nodes or edges as augmentation, and a queue of negatives to learn 64-dimensional structural representations before supervised prediction (Wang et al., 6 Jul 2025).

Training uses an 8:1:1 split for ionic-liquid, polymer, and crystal datasets (Wang et al., 6 Jul 2025). Reported test-set performance spans $\rightarrow$7 to $\rightarrow$8, with representative values including $\rightarrow$9 and MAE $\rightarrow$0 mS/cm for ionic-liquid conductivity, $\rightarrow$1 and MAE $\rightarrow$2 V for ionic-liquid electrochemical window, $\rightarrow$3 and MAE $\rightarrow$4 K for polymer $\rightarrow$5, $\rightarrow$6 and MAE $\rightarrow$7 V for crystal average voltage, and $\rightarrow$8 with MAE $\rightarrow$9 mAh/g for Transformer-based battery C-rate capacity prediction (Wang et al., 6 Jul 2025).

Representative platform predictions illustrate the intended usage. For the ionic liquid C$\rightarrow$0mimFSI, CEMP predicts ECW $\rightarrow$1 V, $\rightarrow$2 K, and $\rightarrow$3 mS/cm; for cathode P2-NaNiO$\rightarrow$4, it predicts $\rightarrow$5 V, $\rightarrow$6 mAh/g, and $\rightarrow$7 mWh/g; and for polymer PPO, it predicts $\rightarrow$8 K, $\rightarrow$9 K, $\rightarrow$0 MPa, and $\rightarrow$1 (Wang et al., 6 Jul 2025). The BMS module further predicts full discharge curves across cycles and C-rates, rather than only scalar summaries (Wang et al., 6 Jul 2025).

These models operationalize rapid screening, structure–property analysis, and effective multi-objective filtering even though the current user interface does not expose explicit Bayesian optimization or genetic algorithms (Wang et al., 6 Jul 2025). In that respect, CEMP differs from autonomous laboratories such as Ada, where model-based optimization is directly embedded in experiment planning via Phoenics (MacLeod et al., 2019). Yet the underlying logic is compatible: CEMP supplies the prediction, simulation, and data infrastructure on top of which a more explicit optimizer could operate.

6. Relationship to self-driving laboratories and MAPs

CEMP should be understood in the broader context of materials acceleration platforms and self-driving laboratories rather than as an isolated database project. The “Ada” laboratory demonstrated a modular robotic platform for thin-film materials in which ChemOS and Phoenics iteratively proposed experiments, a SCARA robot performed liquid handling, spin coating, annealing, imaging, spectroscopy, and four-point-probe measurements, and pseudomobility was computed and sent back to the optimizer in campaigns of 35 samples completed in less than 30 h (MacLeod et al., 2019). That system established a concrete architecture for integrated planning, execution, measurement, and learning.

MAP design studies for heterogeneous CO$\rightarrow$2 photo(thermal)catalysis generalized this approach into a broader blueprint consisting of initialization, AI control, experimentation, characterization, and database feedback, organized around design descriptors, performance descriptors, and multi-scale optimization spanning materials, reactors, and process economics (Wang et al., 2023). They also argued that MAPs must integrate materials, devices, and scale-up descriptors rather than optimizing powders in isolation (Wang et al., 2023).

CEMP shares the same structural logic but currently emphasizes digital unification of database, prediction, and online simulation rather than full experimental autonomy (Wang et al., 6 Jul 2025). This suggests a division of roles: self-driving labs generate high-quality, closed-loop experimental data; platforms such as CEMP standardize, store, predict, and validate; and MAP frameworks provide the systems-engineering perspective required to connect material screening with device and process objectives (MacLeod et al., 2019, Wang et al., 2023). A plausible implication is that CEMP can serve as the digital substrate for federated MAPs, especially if future programmatic access and intelligent-agent orchestration are realized.

This interpretation is reinforced by CEMP’s roadmap, which includes ML-accelerated computation, a Python SDK termed CEMP-Py, intelligent agents or LLM-driven computation for automated job generation and workflow planning, and cloud-native hybrid CPU/GPU infrastructure for elastic scaling and low-latency scheduling (Wang et al., 6 Jul 2025). Those developments move the platform conceptually toward the same closed-loop, self-driving paradigm exemplified experimentally by Ada (MacLeod et al., 2019).

7. Applications, exemplar material entries, and future development

CEMP’s immediate application space includes rapid screening of ionic liquids for low $\rightarrow$3, high ECW, and high conductivity; polymer-electrolyte and separator design using $\rightarrow$4, modulus, dielectric constant, and related properties; crystal-electrode screening using voltage, capacity, and specific energy; and battery-device analysis through Transformer-based C-rate curve prediction (Wang et al., 6 Jul 2025). The platform is explicitly positioned as domain-specific but cross-material, enabling integrated clean-energy workflows rather than crystal-only informatics (Wang et al., 6 Jul 2025).

Its design is especially suitable for multi-domain entries that couple physical descriptors across scales. Janus chromium monolayers exemplify how one material family can simultaneously carry dipole moments, work functions, heterostructure doping behavior, HER descriptors, elastic constants, and magnetic/electronic classification, making them searchable as HER catalysts, low-work-function electrodes, spintronic materials, and vertical p–n junction spacers within one platform record (Li et al., 2022). T-carbon similarly illustrates how a single crystal entry can span structural, electronic, thermoelectric, hydrogen-storage, and battery-anode metadata (Qin et al., 2019). Ce-doped graphene ORR catalysts further show how CEMP can store single-atom catalyst motifs with formation energies and pathway-specific overpotentials for 2e$\rightarrow$5 and 4e$\rightarrow$6 oxygen reduction (Lucchetti et al., 9 May 2025). Process-centric records, such as emulsion-templated inorganic/inorganic composites with tunable porosity, shell thickness, magnetic alignment, and fracture behavior, indicate that microstructure-processing-property datasets can also be incorporated when relevant to solid oxide fuel cells, solid-state batteries, catalytic supports, or structural energy components (Jiang et al., 2024).

The platform nonetheless has clear limitations. Experimental data remain relatively sparse at about 6,000 records compared with computed and ML-generated data; polymer coverage is mainly homopolymers; crystal coverage is currently inherited largely from Materials Project; QC and MD resources are limited to three concurrent tasks; and explicit per-sample uncertainty quantification is not yet fully implemented (Wang et al., 6 Jul 2025). More generally, extrapolation outside current chemical domains remains insufficiently characterized (Wang et al., 6 Jul 2025). These constraints echo broader MAP concerns about data sparsity, heterogeneity, limited automation of key characterization methods, and the need for physics-informed interpretability (Wang et al., 2023).

The longer-term significance of CEMP lies in its attempt to make clean-energy materials discovery executable through a single web-native environment. It does not merely aggregate records; it links structured data, task-specific ML, and online computation in a way that supports rapid screening, structure–property analysis, and iterative validation (Wang et al., 6 Jul 2025). Within the ecosystem of clean-energy research infrastructures, it therefore occupies an intermediate but strategically important position between static databases, autonomous experimental labs, and full multi-scale MAPs (MacLeod et al., 2019, Wang et al., 2023).